尾矿库是用于储存矿石废料的人工构筑物，也是高势能的人造泥石流风险源。不当的监管会导致溃坝、污染物泄漏等灾难性事件发生，对周围环境和居民生命财产安全造成严重威胁。干滩长度作为评估尾矿库安全稳定性的关键指标之一，是指库坝顶到库内水边缘的最短水平距离，其长度过短将严重影响尾矿库的安全。本文利用深度学习方法分割干滩水面，根据图像分割结果求出干滩水面分界线像素坐标，并通过推导得到像素坐标与实际干滩长度的转换公式，进而求得干滩长度。本文具体研究工作如下：

（1）针对尾矿库干滩图像样本数量有限以及因雾气和下雨导致图像对比度和色彩减弱等问题，通过现场拍照和视频截取方式获取尾矿库干滩图像，在此基础上运用暗通道先验法去雾和深度学习方法去雨进行图像优化，利用结构相似度评价指标确保图像不失真，基于现有样本通过多种数据增强方式提高尾矿库干滩样本数量和多样性，从而提升尾矿库干滩分割模型鲁棒性和分割精度。

（2）针对尾矿库干滩水面分界线难以提取、边界线提取粗糙等问题，选取FCN、U-Net、PSPNet、DeepLabV3+四种经典深度学习分割模型，在统一数据集上进行训练和测试，通过对比分析，DeepLabV3+模型在精度指标方面表现最好，其中mIoU值分别比FCN、PSPNet、U-Net高7.27%、5.35%、1.93%，在mPA上DeepLabV3+比FCN、PSPNet、U-Net分别高8.49%、4.33%、2.78%。但DeepLabV3+模型的高精度伴随着更大参数量和更长的推理时间，使其在效率指标上表现较差。在此基础上对DeepLabV3+模型改进，通过替换主干网络以提升语义分割运算效率，降低模型的参数量和推理时间，并调整ASPP结构增强模型对图像特征的捕获能力，引入注意力机制进一步加强对干滩水面分界线重要特征的识别。实验结果表明：改进后的模型mPA提升5.93%，mIOU提升4.89%，参数量从37.12MB减少至16MB，降低了一半以上，推理时间减少了47.93%。改进DeepLabV3+模型更适用于尾矿库干滩水面分界线的提取。

（3）针对传统干滩长度测量方法存在工作量大、时效性低和误差较大等问题，首先，利用陕西某矿区摄像头获取所需分割图像，选用张正友标定法对相机进行标定，以减少相机畸变对长度测量的影响。然后，通过布设控制点采集干滩坝顶和水线边界的坐标信息构建干滩长度的测量模型。最后，应用改进后的DeepLabV3+模型对尾矿库干滩图像进行分割，提取干滩水面边界线的像素值坐标，得到干滩长度值。实验结果表明：干滩长度测量绝对误差小于5m，相对误差小于2%，在陕西省某尾矿库验证该方法满足干滩长度测量实际需求。

Tailings dam is a man-made structure created for the storage of waste materials from ore processing, and it also represents a high-potential source for man-made debris flows. Improper management can lead to catastrophic events such as dam failures and pollutant leaks, posing serious threats to the surrounding environment and the safety of residents' lives and property. The length of the dry beach, as one of the key indicators for assessing the safety and stability of tailings dams, refers to the shortest horizontal distance from the top of the dam to the water's edge inside the dam. A length that is too short can severely compromise the safety of the tailings dam. This paper utilizes deep learning methods to segment the dry beach surface. Based on the results of the image segmentation, we calculate the pixel coordinates of the boundary line between the dry beach and the water surface, and from this, we derive a formula to convert pixel coordinates into actual lengths of the dry beach. The specific research work of this paper is as follows:

(1) Addressing the limited number of tailings dam dry beach image samples and issues such as reduced image contrast and color due to fog and rain, this study acquires dry beach images of tailings dams through on-site photography and video capture. Based on this, the dark channel prior method is used for dehazing and deep learning methods for deraining to optimize images, with the structural similarity index used to ensure image fidelity. Data augmentation methods are applied to existing samples to enhance the number and diversity of tailings dam dry beach samples, improving the robustness and segmentation accuracy of the model.

(2) To address the difficulty of extracting the boundary line between the dry beach and water surface and the coarseness of the boundary line extraction, four classic deep learning segmentation models—FCN, U-Net, PSPNet, and DeepLabV3+—are selected. These models are trained and tested on a unified dataset. Comparative analysis shows that the DeepLabV3+ model performs best in terms of accuracy metrics, with mIoU values respectively 7.27%, 5.35%, and 1.93% higher than FCN, PSPNet, and U-Net, and mPA values 8.49%, 4.33%, and 2.78% higher, respectively. However, the high accuracy of the DeepLabV3+ model comes with a larger number of parameters and longer inference time, making it less efficient. Based on this, the DeepLabV3+ model is improved by replacing the backbone network to improve semantic segmentation operation efficiency, reducing the number of model parameters and inference time, adjusting the ASPP structure to enhance the model's capability of capturing image features, and introducing an attention mechanism to further strengthen the recognition of important features of the dry beach and water surface boundary line. Experimental results show that the improved model increases mPA by 5.93%, mIoU by 4.89%, the parameter size is reduced from 37.12MB to 16MB—more than halved—and inference time is decreased by 47.93%. The improved DeepLabV3+ model is more suitable for the extraction of the boundary line between the dry beach and water surface of tailings dams.

(3) To address the problems of large workload, low timeliness, and large errors associated with traditional methods of measuring the length of the dry beach, this study first acquires the required segmentation images using cameras in a mining area in Shaanxi. The camera is calibrated using the Zhang-Zhengyou calibration method to reduce the impact of camera distortion on length measurement. Control points are then laid out to collect coordinate information for constructing a model to measure the length of the dry beach. Finally, the improved DeepLabV3+ model is applied to segment images of the tailings dam dry beach, extracting pixel value coordinates of the dry beach and water surface boundary line to obtain the dry beach length. Experimental results indicate that the absolute error of dry beach length measurement is less than 5 meters, and the relative error is less than 2%, verifying that this method meets the actual requirements for measuring the length of the dry beach in a tailings dam in Shaanxi Province.